Multi-Source Transfer Learning for Deep Model-Based Reinforcement Learning

Sasso, Remo (2021) Multi-Source Transfer Learning for Deep Model-Based Reinforcement Learning. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2021_SassoR.pdf.pdf Download (3MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (121kB)

Abstract

Reinforcement learning (RL) is well known for requiring large amounts of data for agents to learn tasks. Recent progress in model-based RL (MBRL) allows agents to be much more data-efficient, as it enables agents to learn behaviors in imagination by leveraging an internal World Model of the environment. Improved data-efficiency can also be achieved by reusing knowledge from previously learned tasks, but transfer learning is still an emerging topic for RL. In this research, we propose and investigate novel transfer learning approaches for deep MBRL. Moreover, rather than transferring knowledge from a single source, this work focuses on multi-source transfer learning techniques. First, we propose transferring knowledge of agents that were trained on multiple tasks simultaneously. In addition, where common transfer learning techniques use an all-or-nothing approach for transferral of neural network layers, we present fractional transfer learning as an alternative approach. Next, we introduce meta-model transfer learning, which is a technique that allows the combination and transferral of knowledge for multiple individual sources in a universal feature space. Finally, we present latent task classification, which enables agents to classify previous tasks and detect novel tasks in latent space. Experimental results show he proposed transfer learning methods can lead to significantly faster and/or better learning performances, and verify the ability of latent task classification.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Wiering, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	12 Jul 2021 08:29
Last Modified:	12 Jul 2021 08:29
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/25168

Actions (login required)

View Item